AI news, 20 September: Trump plans ‘AI Force’ and new czar

Trump announces ‘AI Force’ and AI czar. U.S. President Donald Trump said on Saturday via Truth Social that he will form an “AI Force,” modeled on the Space Force he created in his first term, and soon name an AI czar, according to reports from the Washington Post, Reuters, CNN and the New York Times. He framed the move as supporting industry growth without new constraints, dismissing many safety concerns as a “hoax” while saying existing criminal and civil laws would address “BAD” behavior. Details on structure, powers or whether it would be military or civilian remain unclear. The announcement comes amid industry calls to slow advanced AI development and ahead of expected talks with China’s Xi Jinping.

Anthropic weighs new model before IPO. According to a Reuters exclusive citing sources, Anthropic is considering releasing a new frontier model to counter momentum from OpenAI’s GPT-6 Astra, which has captured notable enterprise spending share. The discussions occur even after CEO Dario Amodei urged the industry to pace capabilities for safety reasons. Reports also indicate Anthropic is targeting a November IPO, with annualized revenue projections climbing toward or beyond $100 billion this year, up sharply from earlier figures.

Anthropic embeds Accenture safety evaluators. Anthropic announced a partnership with Accenture (via its Faculty AI unit) to place third-party evaluators inside the company with employee-like access. They will red-team models, assess alignment and test safeguards. Each side expects to invest at least $1 billion over five years; Anthropic will fund the work directly for now. The company called it a step toward the embedded-evaluator model outlined in Amodei’s recent slowdown proposal and said talks continue with nonprofit groups such as METR.

Zero-click flaw hits major AI coding agents. Security researchers at AIR disclosed Plugin4Shell, a zero-click remote code execution vulnerability affecting Claude Code, OpenAI Codex, GitHub Copilot and Gemini CLI. It bypasses SHA pinning on marketplace plugins, potentially allowing malicious code via auto-updates. Anthropic and OpenAI have issued patches; Microsoft had not fully patched Copilot at disclosure, and Google deprecated Gemini CLI rather than fixing it, according to Help Net Security and related reports.

AI error nearly prompted U.S. action on Chinese ship. CNN reported that a U.S. Special Operations Command analyst used a chatbot that produced an “entirely false” intelligence assessment claiming a Chinese vessel in the Middle East carried nuclear weapons components. Military plans to intercept and board advanced far enough that planes were aloft before the error was caught; one source said it “almost started a war.” The episode, from earlier this year, underscores risks of AI in high-stakes intelligence work.

Alibaba open-sources expert-level medical CT model. Alibaba’s DAMO Academy released Damo Radar (also called RADAR), a vision-language model that analyzes abdominal CT scans for nearly 150 conditions across 18 organs. Published in Science and open-sourced, it reportedly achieved strong accuracy (average AUC around 0.913 on large real-world tests) and outperformed most of 26 radiologists in head-to-head evaluation, while helping doctors cut missed diagnoses and reading time when used as an aid.